--- datasets: - jan-hq/instruction-speech-v1 language: - en license: apache-2.0 tags: - sound language model --- ## Model Details We have developed and released the family [Jan-Llama3](https://huggingface.co/collections/jan-hq/jan-llama3-668e4dad446c8736208dca4f). This family is natively understanding audio and text input. We continue to expand [Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) with sound understanding capabilities by leveraging 700M tokens [Instruction Speech v1](https://huggingface.co/datasets/jan-hq/instruction-speech-v1) dataset. **Model developers** Homebrew Research. **Input** Text and sound. **Output** Text. **Model Architecture** Llama-3. **Language(s):** English. ## Intended Use **Intended Use Cases** This family is primarily intended for research applications. This version aims to further improve the LLM on sound understanding capabilities. **Out-of-scope** The use of Llama-3-Sound in any manner that violates applicable laws or regulations is strictly prohibited. ## How to Get Started with the Model First, we need to convert the audio file to sound tokens ```python import torch import torchaudio from encodec import EncodecModel from encodec.utils import convert_audio def audio_to_sound_tokens(audio_path, target_bandwidth=1.5, device="cuda"): # Initialize Encodec model = EncodecModel.encodec_model_24khz() model.set_target_bandwidth(target_bandwidth) model.to(device) # Load and preprocess audio wav, sr = torchaudio.load(audio_path) wav = convert_audio(wav, sr, model.sample_rate, model.channels) wav = wav.unsqueeze(0).to(device) # Encode audio with torch.no_grad(): encoded_frames = model.encode(wav) codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1) # Flatten codes audio_code1, audio_code2 = codes[0][0], codes[0][1] flatten_tokens = torch.stack((audio_code1, audio_code2), dim=1).flatten().tolist() # Convert to sound tokens result = ''.join(f'<|sound_{num}|>' for num in flatten_tokens) return f'<|sound_start|>{result}<|sound_end|>' # Usage sound_tokens = audio_to_sound_tokens("/path/to/your/audio/file") ``` Then, we can inference the model the same as any other LLM. ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline def setup_pipeline(model_path, use_4bit=True): tokenizer = AutoTokenizer.from_pretrained(model_path) model_kwargs = {"device_map": "auto"} if use_4bit: model_kwargs["quantization_config"] = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4", ) model = AutoModelForCausalLM.from_pretrained(model_path, **model_kwargs) return pipeline("text-generation", model=model, tokenizer=tokenizer) def generate_text(pipe, messages, max_new_tokens=64, temperature=0.0, do_sample=False): generation_args = { "max_new_tokens": max_new_tokens, "return_full_text": False, "temperature": temperature, "do_sample": do_sample, } output = pipe(messages, **generation_args) return output[0]['generated_text'] # Usage llm_path = "jan-hq/Jan-Llama3-0708" pipe = setup_pipeline(llm_path, use_4bit=True) messages = [ {"role": "user", "content": sound_tokens}, ] generated_text = generate_text(pipe, messages) print(generated_text) ``` ## Training process **Training Metrics Image**: Below is a snapshot of the training loss curve visualized. ![train_loss_curve/png](https://cdn-uploads.huggingface.co/production/uploads/65713d70f56f9538679e5a56/9bv-kpnqrTxaBhiYrVHN7.png) ### Hardware **GPU Configuration**: Cluster of 8x NVIDIA H100-SXM-80GB. **GPU Usage**: - **Continual Training**: 8 hours. ### Training Arguments | Parameter | Continual Training | |----------------------------|-------------------------| | **Epoch** | 1 | | **Global batch size** | 128 | | **Learning Rate** | 5e-5 | | **Learning Scheduler** | Cosine with warmup | | **Optimizer** | [Adam-mini](https://arxiv.org/abs/2406.16793) | | **Warmup Ratio** | 0.1 | | **Weight Decay** | 0.01 | | **beta1** | 0.9 | | **beta2** | 0.98 | | **epsilon** | 1e-6 | | **Gradient Cliping** | 1.0 | ### Accelerate FSDP Config ``` compute_environment: LOCAL_MACHINE debug: false distributed_type: FSDP downcast_bf16: 'no' enable_cpu_affinity: true fsdp_config: fsdp_activation_checkpointing: true fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP fsdp_backward_prefetch: BACKWARD_PRE fsdp_cpu_ram_efficient_loading: true fsdp_forward_prefetch: false fsdp_offload_params: false fsdp_sharding_strategy: FULL_SHARD fsdp_state_dict_type: SHARDED_STATE_DICT fsdp_sync_module_states: true fsdp_use_orig_params: false machine_rank: 0 main_training_function: main mixed_precision: bf16 num_machines: 1 num_processes: 8 rdzv_backend: static same_network: true tpu_env: [] tpu_use_cluster: false tpu_use_sudo: false use_cpu: false ``` ## Examples 1. Good example: ``` # Question: Does the following review have a positive or negative opinion of the movie? i thought the relationships were wonderful , Options: A). negative; B). positive; # Expected Output: The review has a B). positive opinion about the movie because they think the relationships were wonderful. # Model's Output: The review is positive. ``` 2. Misunderstanding example: ``` # Question: Translate the following sentence to Russian: 6 Work in JOBS and careers after 40 Russian: # Expected Output: 6 Работать в РАБОТАХ и карьере после 40 Russian (Cyrillic): 6 Работать в РАБОТАХ и карьере после 40 Let me explain this to you like you're five years old: This sentence means that people can still work and have careers, even if they're older than 40 years old. The words are in a different language, called Russian. # Model's Output: В случае, когда вы хотите перевести фразу "The first step is to identify the language of the text." на русский язык, вы можете использовать следующую фразу: "Первый шаг - определить язык текста." ``` 3. Off-tracked example: ``` # Quesion: Answer the following question: - The roots of the growing plants start to break up the rock - The plant acids dissolve the rock - The rock is broken down into smaller pieces - Erosion begins. What might be the first step of the process? Answer: # Expected Output: The first step of the process might be: The roots of the growing plants start to break up the rock. # Model's Output: Step 1: Identify the key elements in the premise and hypothesis. Premise: "A man in a blue shirt and a red hat is holding a sign." Key elements: man, blue shirt, red hat, holding, sign Hypothesis: "A man is holding a sign." Key elements: man, holding, sign Step 2: Compare the key elements in the premise and hypothesis. In the premise, we have a man wearing a blue shirt and a red hat, while in the hypothesis, we only have a man. The premise also mentions that the man is holding a sign, which is also mentioned in the hypothesis. Step 3: Determine if the hypothesis is entailed by the premise. Since the hypothesis states that a man is holding a sign, and the premise confirms that a man is holding a sign, we can conclude that the hypothesis is entailed by the premise. The additional information about the man's clothing in the premise does not contradict or negate the ``` Despite being undertrained, the model demonstrates an emerging grasp of sound-text semantics. ## Citation Information **BibTeX:** ``` @article{Llama-3-Sound: Sound Instruction LLM 2024, title={Llama-3-Sound}, author={Homebrew Research}, year=2024, month=July}, url={https://huggingface.co/jan-hq/Jan-Llama3-0708} ``` ## Acknowledgement - **[WhisperSpeech](https://github.com/collabora/WhisperSpeech)** - **[Encodec](https://github.com/facebookresearch/encodec)** - **[Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct)**